Research On Traffic Sign Recognition Based On Neural Network

With the development of our country’s economy and information technology,the penetration rate of automobiles has greatly increased,but on the other hand,various traffic accidents have also followed.In order to cope with the frequent occurrence of traffic accidents,advanced driver assistance systems(TSR)came into being,which can help drivers make correct driving operations,thereby effectively avoiding traffic accidents.In TSR,the accurate recognition of traffic signs is a core issue,which is of great research significance.In order to further to improve the accuracy and calculation efficiency of traffic signs recognition,the main work of this paper are as follows :First of all,In order to solve the problem of unbalanced sample size and excessive noise in the GTSRB data set in the traffic sign recognition algorithm,this paper performs sample equalization and image enhancement processing on the data set.First,the method of translation and flip is used to expand each type of samples of about 2500,so as to achieve sample balance.In addition,on the equalized data set,the improved algorithm of Scharr edge detection after histogram equalization and Gaussian blur is used for image enhancement.Experiment with the data set after image enhancement,and the result proves that compared with the original method,the accuracy is effectively improved.Secondly,In view of the large amount of calculation of the deep learning network model,this paper proposes a new type of network that combines the existing convolutional network VGG16 with the hetero-core convolutional network HETCONV,and makes the model lightweight for the original model..That is,the HETCONV layer is used to replace all the layers of the convolutional network VGG16 except the first layer,which reduces the size of the convolution kernel.In the subsequent simulation experiments,the improved convolutional network was compared with the original convolutional network.The experimental results show that compared with the existing convolution algorithm,the new network HET-VGG16 convolution model can greatly reduce the amount of calculation and reduce training time.Once again,In order to prevent the network model from over-fitting,this paper further optimizes the network model.The improved convolutional network is divided into5 convolutional structures,and each convolutional structure is added with a batch normalization regularization layer and a dropout layer.The dropout dropout rate of the first four convolutional structures is set to 0.2,and the dropout dropout rate of the last convolutional structure is set to 0.5.At the same time,comparative experiments were carried out on the selection of various network parameters,and the optimal model was obtained.